Adaptively Weighted L2-Minimization in Predictive Speech Coding

Current narrow-band speech coding algorithms (for transmission rates of 2400-4800 bits per second) typically excite linear filters with impulse trains to model voiced speech. The excitation function that would reproduce the speech exactly is the prediction residual; however, the usual selection of filter coefficients does not produce the most pulse-like prediction residuals. Thus other choices for filters offer an opportunity to improve the quality of narrow-band coding. The strategy of this paper is to minimize a dynamically weighted prediction error, to allow the largest values of the prediction residuals to be unconstrained and thus make the residuals more pulse-like. The idea was tested on voiced speech with over 300 predictive models, which contained quadratic and cubic terms as well as linear. Models with fewer than eight terms were not enhanced. The idea worked well with other models, particularly those with 8 to 11 terms.